Publication number | US7487425 B1 |

Publication type | Grant |

Application number | US 08/062,894 |

Publication date | Feb 3, 2009 |

Filing date | May 17, 1993 |

Priority date | Mar 6, 1989 |

Fee status | Lapsed |

Also published as | EP0386506A2, EP0386506A3, US5600659 |

Publication number | 062894, 08062894, US 7487425 B1, US 7487425B1, US-B1-7487425, US7487425 B1, US7487425B1 |

Inventors | Chin-Long Chen |

Original Assignee | International Business Machines Corporation |

Export Citation | BiBTeX, EndNote, RefMan |

Patent Citations (4), Non-Patent Citations (10), Referenced by (14), Classifications (7), Legal Events (3) | |

External Links: USPTO, USPTO Assignment, Espacenet | |

US 7487425 B1

Abstract

Byte or symbol organized linear block codes are optimized in terms of reducing the number of ones in their parity check matrices by means of symbol column transformations carried out by multiplication by non-singular matrices. Each optimized symbol column preferably, and probably necessarily, includes a submatrix which is the identity matrix which contributes to low weight check matrices and also to simplified decoding procedures and apparatus. Since circuit cost and layout area are proportional to the number of Exclusive-OR gates which is determined by the number of ones in the check matrix, it is seen that the reduction procedures carried out in accordance with the present invention solve significant problems that are particularly applicable in the utilization of byte organized semiconductor memory systems. Reduced weight coding systems are also generated in accordance with weight reducing procedures used in conjunction with modified Reed Solomon codes. The decoding and encoding methods and apparatus are extendable to the inclusion of any number of check symbols.

Claims(19)

1. A method for encoding input digital electrical signals into blocks of n bits and subblocks of at most b symbol bits, said method comprising the steps of:

generating check bit signals from said input digital electrical signals so that said check bit signals satisfy parity check conditions determined by a binary parity check matrix derivable from a matrix of the form H=[H_{1 }H_{2 }. . . H_{N}] wherein each symbol column, H_{j}, j=1, 2, . . . , N, is a column of M disjoint binary submatrices having b rows and at most b columns and wherein at least one of said N symbol columns has been modified to provide a minimal number of ones in said symbol column; and

combining said check bit signals with said input signals so as to produce a block of n signals having redundancy for improved error correction and detection.

2. The method of claim 1 in which all of said N symbol columns possess a minimal number of ones.

3. The method of claim 1 in which M=3.

4. The method of claim 1 in which each column of submatrices includes at least one identity matrix.

5. The method of claim 1 in which each submatrix is either a matrix of zeroes or a power of a matrix T which is the companion matrix associated with a primitive polynomial over GF(2).

6. The method of claim 5 in which said primitive polynomial is p(x)=1+x^{2}+x^{5}.

7. The method of claim 5 in which M=3 and in which, in at least one symbol column, each submatrix is distinct and selected from the group consisting of the matrices I, T^{q }and T^{−q }being an integer.

8. The method of claim 3 in which M=3 and in which in each symbol column each submatrix is distinct and selected from the group consisting of the matrices I, T^{q }and T^{−q }being an integer.

9. A method for decoding electrical signals representing a linear binary code occurring in blocks of n bits and subblocks of at most b symbol bits, said electrical signals having been encoded subject to parity check conditions determined by a parity check matrix derivable from a matrix H of the form [H_{1 }H_{2 }. . . H_{N}], wherein each symbol column H_{j}, j=1, 2, . . . , N, is a column of M disjoint binary submatrices H_{ij }and wherein, for at least one i, H_{ij }includes an identity matrix with b rows and at most b columns, said decoding method comprising the steps of:

determining a syndrome vector S for a received binary vector y, said syndrome S having M subvectors S_{1}, S_{2}, . . . , S_{M};

determining, for at least one symbol position j, the (M−1) matrix-vector products H_{ij}·S_{k}, for i not equal to k, where k denotes the matrix row of H_{j }which includes said identity matrix;

comparing said (M−1) matrix-vector products H_{ij}·S_{k }with corresponding syndrome vectors S_{i }to determine the presence of correctable symbol errors; and

correcting at least said j^{th }symbol using the vector pattern S_{k }upon indication of correctable symbol errors from said comparison.

10. The decoding method of claim 9 which includes correcting all of said symbols.

11. The method of claim 9 in which there is a minimal number of ones amongst the binary elements in H_{ij}, H_{2j}, . . . , H_{mj }for some j greater than 1 and less than or equal to M.

12. The method of claim 9 in which each H_{ij }is either a null matrix or a power of a matrix T which is the companion matrix for a primitive polynomial over GF(2).

13. The method of claim 9 in which M=3.

14. The method of claim 9 in which said primitive polynomial is p(x)=1+x^{2}+x^{5}.

15. The method of claim 9 in which the determination of H_{ij}·S_{k }and said comparison is carried out in a single layer of Exclusive-OR gates.

16. An apparatus for decoding electrical signals representing a linear binary code occurring in a block of n bits and the subblocks of at most b symbol bits, said electrical signals having been encoded subject to parity conditions determined by a binary parity check matrix derivable from a matrix H of the form [H_{1 }H_{2 }. . . H_{N}], wherein each symbol column H_{j}, j=1, 2, . . . , N, is a column of M disjoint binary submatrices, H_{ij}, i=1, 2, . . . , M one of which is an identity matrix with b rows and at most b columns, said apparatus comprising:

means for determining a syndrome vector S for a received binary vector y, said syndrome S having M subvectors S_{i}, S_{2}, . . . , S_{M};

means for determining, for at least one symbol position j, the (M−1) matrix-vector products H_{ij}·S_{k }for i not equal to k, where k denotes the matrix row of H_{j }which includes said identity matrix;

means for comparing said (M−1) matrix-vector products with corresponding syndrome subvectors S_{i }to determine the presence of correctable errors; and

means for correcting at least said j^{th }symbol using the vector pattern S_{k }upon indication of correctable symbol errors from said comparison means.

17. The apparatus of claim 16 in which said means for determining said matrix-vector product and said comparison means are implemented as a single layer of Exclusive-OR gates.

18. The apparatus of claim 16 in which M=3.

19. A method for generating check bit signals from digital input signals to form a coded signal having blocks of n bits and subblocks of at most b symbol bits, said method comprising the step of:

generating, with a parity check circuit, check bit signals from said input digital electrical signals so that said check bit signals satisfy parity check conditions determined by a binary parity check matrix derivable from a matrix of the form H=[H_{1 }H_{2 }. . . H_{N}], wherein each symbol column, H_{j}, j=1, 2, . . . , N, is a column of M disjoint binary submatrices having b rows and at most b columns and wherein at least one of said N symbol columns has been modified to provide a minimal number of ones in said symbol column.

Description

This is a continuation of application(s) Ser. No. 07/318,983 filed on Mar. 6, 1989, now abandoned.

The present invention is generally directed to the encoding and decoding of binary data so as to employ error correction coding (ECC) methods for symbol error correction and detection. Apparatus for decoding such codes is also provided and implemented in a fashion which exploits code structure to provide encoders and decoders which minimize circuit cost, particularly with respect to the number of Exclusive-OR gates needed. More particularly, an embodiment employing three check symbols is presented.

The utilization of error correction and detection codes in electronic data processing and transmission systems is becoming more and more important for several reasons. In particular, increased problem complexity and security concerns require ever increasing levels of reliability in data transmission. Furthermore, the increased use of high density, very large scale integrated (VLSI) circuit chips for use in memory systems has increased the potential for the occurrence of soft errors such as those induced by alpha particle background radiation effects. Furthermore, the increased use of integrated circuit chips has led to memory organizations in which each output memory word is derived from multiple bit patterns received from a plurality of different circuit chips. Accordingly, it has become more desirable to be able to protect memory system integrity against the occurrence of chip failures since such failures produce multi-bit errors which are associated with a single chip. It is thus seen that it is desirable to employ error correction coding systems which take advantage of the memory system organization itself, particularly with respect to minimizing the probability of error. For this reason, it is desirable to employ codes which are based upon symbols or bytes of data. (It is noted that, as used herein, the term “byte” does not necessarily refer to an 8-bit block of binary data but rather is more generic.)

As noted, error correcting codes have been applied to semiconductor memory systems to increase reliability, reduce service costs and to maintain data integrity. In particular, single error correcting and double error detecting (SEC-DED) codes have been successfully applied to many computer memory systems. These codes have become an integral part of the memory design for medium and large systems throughout large portions of the computer industry.

However, the error control effectiveness of an error correction code depends on how the memory chips are organized with respect to the code. In the case of single error correcting and double error detecting codes, the one bit per chip organization is the most effective design. In this organization, each bit of a code word is stored in a different chip, thus any type of failure in a chip can corrupt at most one bit of the code word. As long as the errors do not line up in the same code word, multiple errors in the memory are correctable in this scheme.

As the trend in chip design has continued toward higher and higher density, it has become more difficult to design one bit per chip types of memory organizations because of the system granularity problem. For example, the system capacity has to be at least four megabits if one megabit chips are used to design a memory with a 32 bit data path. However, in a b bit per chip memory, a chip failure may result in from 1 to b bit errors, depending upon the type of failure. A failure could be a cell failure, a word line failure, a bit line failure, a partial chip failure or a total chip failure. With a maintenance strategy that allows correctable errors in the memory to accumulate, SEC-DED codes are not particularly suitable for b bit per chip memories. Since multiple bit errors are not correctable by SEC-DED codes, the uncorrectable error rates can be high if the distribution of chip failure types is skewed to those types that result in multiple bit errors.

A more serious problem is the loss of data integrity due to the miscorrection of some multiple errors. It is therefore seen that for byte organized memory systems, it is desirable to employ single byte error correcting and double byte error detecting (SBC-DBD) codes in b bit per chip memories. With such codes, errors generated by a single chip failure are always correctable and errors generated by a double chip failure are always detectable. Thus, the uncorrectable error rate can be kept low and the data integrity can be maintained. Discussions concerning codes that are suitable for this purpose can be found in the article “Error-Correcting Codes for Byte-Organized Memory Systems”, by Chin-Long Chen in Volume IT-32, No. 2 of the IEEE Transactions on Information Theory (March 1986).

In the construction of error correcting codes contrary criteria are often present. The first criteria relates to the error correcting capabilities of the code itself. The second criteria relates to the ease of encoding and more particularly, of decoding the information received. More complex codes often require more complicated circuitry for decoding. With the constraints of electronic circuit chip “real estate” being at a premium with respect to circuit design, it is seen that it is greatly desirable to be able to implement both encoding and decoding circuits utilizing as few circuit gates as possible with each gate having a minimum number of attached signal lines. Accordingly, one of the objects of the present invention is the construction of minimum cost encoders and decoders without the corresponding sacrifice of correction and detection capabilities.

In particular, one of the classes of codes which is particularly applicable to the present invention is a code employing three check symbols. Such a code can be described as an extended Reed Solomon code having a parity check matrix of the form:

In Equation 1 above, H is the parity check matrix, a well known concept in error correction coding. I represents the identity matrix and O represents a matrix all of whose elements are zero. In general, T is a matrix with b rows and b columns and represents the companion matrix of a b^{th }degree primitive polynomial over the field GF(2). In one particular embodiment of this form of Reed Solomon code, the integer a is taken to be zero. Further information concerning such codes is found in the aforementioned article by the present inventor.

In accordance with a preferred embodiment of the present invention, a method for encoding binary data which occurs in a block of n bits and subblocks of at most b symbol bits, comprises subjecting the n bits of data to parity check conditions determined by a binary parity check matrix which is either equal to or derivable from a matrix of the form H=[H_{1 }H_{2 }. . . H_{N}], where each symbol column H_{j }is a column of M disjoint binary submatrices H_{ij }having b rows and at most b columns and wherein at least one of the N symbol columns possesses a minimal number of ones. In particular, the present inventor has realized that the submatrices in each symbol column may be multiplied by non-singular binary matrices without affecting error correction or detection properties. Nonetheless at the same time the resultant code exhibits a reduction in the number of ones in the parity check matrix. It is seen that this is a desirable goal since the number of 2-way Exclusive-OR gates required to generate the i^{th }check bit is equal to the number of ones in the i^{th }row of the binary matrix H less one. Thus, in the situation in which each symbol possesses b bits, the total number of Exclusive-OR gates required to generate all check bits is equal to the number of ones in the binary matrix H−3b (for M=3), M being the number of check symbols. In addition, the number of ones in a symbol-column of H is proportional to the number of Exclusive-OR gates required to decode syndrome errors. Thus, to reduce error correction encoding and decoding hardware, it is seen it is desirable to reduce the number of ones in each symbol-column, H_{j}.

Thus, in accordance with the present invention it is possible to manipulate symbol columns of H through the utilization of multiplication by non-singular matrices which thus create equivalent codes. While multiplication of the H matrix itself is known to effect the creation of equivalent codes having identical error correction and detection capabilities, this method does not appear to have been applied to symbol columns in the parity check matrices of byte organized codes.

In the process of symbol column manipulation taught by the present invention, it is further desired to carry out non-singular transformations for symbol columns independently of one another in such a way that one of the M disjoint binary submatrices which comprise column H_{j }is an identity matrix. It is not known, but it is suspected, that it is a necessary condition for minimality that at least one of the M disjoint submatrices is in fact an identity matrix. It is clearly seen that this is a desirable goal in that an identity matrix itself possesses a minimal number of ones and cannot in fact possess fewer ones without a reduction in rank. Thus, in accordance with preferred embodiments of the present invention, the symbol columns in the parity check matrices all possess at least one identity submatrix. Furthermore, it is seen that this situation is always possible to create, particularly with respect to the situation involving Reed Solomon codes. The constraint of having an identity matrix present further simplifies the design of decoding circuitry. Although each symbol decoding unit must be designed separately, nonetheless, there is a decided saving in the number of Exclusive-OR gates required, sometimes resulting in a reduction in circuitry by as much as approximately 25%. Furthermore, the separate design requirement is only a factor during code design and does not affect coding or decoding cost or speed.

In accordance with yet another embodiment of the present invention, a method for decoding linear binary codes of the kind indicated above comprises the steps of determining a syndrome vector S for a received binary vector y with the syndrome S being treated as having M subvectors S_{1}, S_{2}, . . . , S_{M}. Then, for at least one symbol position j the (M−1) matrix-vector products H_{ij}·S_{k }(i≠k) are determined. In the above, k denotes the matrix row of H_{j }which is an identity matrix. The value of k depends on j and varies from symbol column to symbol column. The (M−1) vector products are then compared with the corresponding syndrome vectors S_{i }to determine the presence or absence of correctable symbol errors. Lastly, at least the j^{th }symbol is corrected using the vector pattern S_{k }as an error pattern whenever an the indication of a correctable symbol error arises from the comparison operation. A corresponding apparatus for decoding such linear binary codes is also presented herein.

Accordingly, it is an object of the present invention to permit the design of parity check matrices having a minimal weight, that is having a minimal number of 1's in their structure.

It is also an object of the present invention to extend the applicability of at least a subset of Reed Solomon codes.

It is yet another object of the present invention to construct error correcting codes and encoding apparatus which consume minimal amounts of circuitry, particularly as measured with respect to area consumed on VLSI circuit chips.

It is a still further object of the present invention to construct decoding apparatus in which the aforementioned matrix-vector product operations and comparison operations are carried out in a single layer of Exclusive-OR gates.

It is also an object of the present invention to construct codes, encoders and decoding systems which are particularly applicable to byte organized memory systems without in any way impairing the error correction or detection capabilities of such codes.

It is yet another object of the present invention to be able to construct codes which are derivable from minimal weight codes so as to further simplify coder and decoder design and implementation.

Lastly, but not limited hereto, it is an object of the present invention to implement low cost error correcting codes which enhance computer storage integrity, reliability and serviceability.

The subject matter which is regarded as the invention is particularly pointed out and distinctly claimed in the concluding portion of the specification. The invention, however, both as to organization and method of practice, together with the further objects and advantages thereof, may best be understood by reference to the following description taken in connection with the accompanying drawings in which:

_{i};

_{j }of the parity check matrix H;

**1** and A**2**) illustrates the particular circuitry employed to implement the matrix vector product and comparison operations performed for symbol j=1 for the matrix shown in

**1** and B**2**) is similar to

The description which follows first considers the situation in which the number of check symbols in the code, M, is equal to 3. However, in the latter part of the present description, attention is directed to the general case in which M is greater than 3.

In one embodiment, the present invention is directed to a class of binary codes which have been found to be useful because of their error correction properties and because of their ease of decoding. However, the present invention improves upon these codes with respect to both encoding and decoding, particularly when measured against circuit cost as determined by the number of Exclusive-OR gates employed. The present invention is specifically directed to code construction based upon codes having a parity check matrix organized in the form of a plurality of symbol columns. Each of the symbol columns preferably includes submatrices which are powers of a given companion matrix, T, which is the matrix associated with a primitive polynomial over the Galois Field GF(2). For example, if the primitive polynomial p(x) is of degree b and has the following form:

*p*(*x*)=1+*p* _{1} *x+p* _{2} *x+p* _{2} *x* ^{2} *+ . . . +p* _{b−1} *x* ^{b−1} *+x* ^{b}, (2)

then the companion matrix T is of the form given below:

It is noted that in the mathematical field at hand the elements, namely 0 and 1, are their own additive inverses since addition is carried out in a modulo 2 fashion. This fact accounts for the absence of minus signs which would otherwise normally be present in the form given for the T matrix for other fields. Generally, the companion matrix T possesses b rows and b columns. The matrix T is also always non-singular. However, certain code simplifications and reductions, which are more particularly described below, can alter this result, at least formally, when specific binary parity check matrices are derived by column or row deletion. Appropriate parity check matrices are developed by computing the various powers of the companion matrix T and inserting such matrices in the positions shown, for example as in Equation (1) above. The result is a binary parity check matrix with specific symbol error correction properties in which each symbol is represented by at most b bits.

In one particular embodiment of the present invention, it is convenient to consider parity check matrices in which the number M of check symbols is 3. Since the submatrices, T, each generally include b rows and b columns and because of the structure of the code and the primitive nature of the polynomial p(x), there are at most 2^{b}−2 such matrices (not counting the matrix T^{0 }which is equal to the identity matrix). Thus it is useful, at least initially, to consider parity check matrices having the following form:

In the above m (which is unrelated to M) is less than or equal to 2^{b}−2. Also in the above I represents in general, an identity matrix with b rows and b columns the matrix O is a b×b matrix with all zero elements. The code defined by Equation (4) requires 3b binary check bits whose positions correspond to the last 3b binary columns of H (for the matrix shown). It is further noted that the matrix above is of the form H=[H_{1 }H_{2 }. . . H_{N}], where N is the number of code symbols, that is the number of blocks of b bits in a code word. In this respect, as long as each symbol is represented by b bits, the number of bits n in a code word is Nb. In the matrix H above, each of the N columns represents a binary b column matrix. These columns are referred to herein as symbol columns. Also, the weight of a binary column is the number of 1's in the column and the weight of a symbol column is the number of 1's in the b constituent binary columns of the symbol column.

One of the observations that is key to an understanding of the present invention is the fact that each symbol column H_{j }can multiply a b×b non-singular matrix which has the effect of forming a different code by scrambling the bits in a symbol column in an invertible linear fashion. The code characteristics however are not changed. This manipulation is possible since the code is being treated as a symbol code having N separate b bit blocks. The individual b bit blocks are independently manipulable in this fashion. In fact, the manipulation of the symbol columns of H by multiplication by non-singular matrices is not dependent on the symbol columns of H having any specific form. As indicated, it is applicable to any parity check matrix for byte organized codes.

The motivation for this matrix multiplication arises from the fact that the number of two-way Exclusive-OR gates required to generate i^{th }check bit is equal to the number of 1's in the i^{th }row of the binary matrix of H, less 1. Thus the total number of Exclusive-OR gates required to generate all check bits is equal to the number of 1's in the binary matrix of H, less 3b. It is thus seen that to reduce error correction coding hardware, it is desirable to reduce the number of 1's in each symbol column and in general to reduce the number of 1's in the binary matrix H itself. In general, the parity check matrices H shown in Equations (1) or (4) are not optimal in this sense.

In general, each symbol column is multiplied by all b×b non-singular binary matrices and the resulting symbol column having the lowest number of 1's is thereafter employed as the symbol column of choice. Thus, in the present invention symbol columns are employed which exhibit optimality in the sense that each symbol column possesses a minimum number of 1's. Such symbol columns and the parity check matrices comprised of such symbol columns are optimal in the sense that they exhibit minimal hardware requirements.

It is suspected, though it is not certain, that general optimization in this fashion produces symbol columns which exhibit an identity submatrix somewhere in the stack of disjoint submatrices that comprise the symbol column. However, in its most general embodiment, the present invention is not specifically limited to this requirement. Nonetheless, for purposes of decoding and for establishing error patterns, it is highly desirable to be able to construct parity check matrices with symbol columns which do include an identity submatrix. As seen below, this problem is particularly easily solved in the case of parity check matrices of the kind illustrated in Equations (1) or (4) above.

In particular, attention is now more specifically directed to optimization methods by parity check matrix modification which can be carried out on matrices of the specific form shown in Equation (4). In particular, it is seen that, apart from the first column and the last three symbol columns, each symbol column includes an identity matrix in the first matrix row and matrices of the form T^{q }and T^{−q }in the other two matrix rows. Because of the structure of the matrix T, it is in itself non-singular and maybe employed to perform the non-singular symbol column transformations indicated above. For example, it is possible to multiply symbol column #**2** in the matrix from Equation (4) by either T or T^{−1 }to form symbol columns having either of the following structures respectively:

or

In each case it is seen that the resultant symbol column H_{2 }includes an identity matrix. Nonetheless, the three possible symbol columns H_{2 }indicated in Equations (4), (5a) and (5b) do not necessarily possess the same weight. Nonetheless, it is possible without changing code characteristics to select any one of these three possibilities as symbol column # 2 for parity check matrix H. This is a key aspect to understanding the operation and effect of the present invention. In particular, the present invention provides a low cost symbol error correction coding scheme which has a minimal weight in at least one and preferably in each symbol column of its parity check matrix. It is further noted that it is not necessary to have an identity matrix as the first non-zero submatrix of a symbol column in the parity check matrix. As long as there is an identity matrix somewhere in the symbol column, the error pattern for the erroneous symbol can always be readily obtained. This observation is key to the construction of minimal weight symbol columns in the parity check matrix and a key to the construction of low cost decoding circuitry. It is also seen that the matrix structure in Equation (4) is particularly applicable to the construction of codes in which an identity matrix is always present in a symbol column.

The above discussion leads to a simple algorithm for the construction of improved parity check matrices. Such matrices produce lower cost encoders and are particularly amenable to the production of lower cost decoders especially in the situation in which an identity matrix is present in each symbol column. More particularly, for each symbol column H_{q }of the form:

where q is an integer, the following procedure may be employed to improve the check matrix H by modification of the symbol columns. In particular, let W_{1 }be the weight of H_{Q}. Next, multiply the matrix T^{−q }by the column H_{Q }and let W_{2 }be the weight of this modified symbol column. Next, multiply the matrix T^{q }by the symbol column H_{Q }and let W_{3 }be its corresponding weight, that is the number of 1's in the resultant matrix. Next, the 3 weights W_{1}, W_{2 }and W_{3 }are compared. If W_{1 }is the smallest, the matrix H_{Q }remains unchanged. If W_{2 }is the smallest, then H_{Q }is replaced by H_{Q}·T^{−q}. Similarly, if W_{3 }is the smallest weight, then the symbol column H_{Q }is replaced by the symbol column H_{Q}·T^{q}. It is noted that these symbol column manipulations are preferably carried out for as many symbol columns as possible. It is also noted that a symbol column substitution is not necessarily made in all instances. That is to say, W_{1 }will not always be greater than W_{2 }and W_{3}.

The above principles are illustrated in the following example in which b=5, m=14 and p(x)=1+x^{2}+x^{5}. In this case, the companion matrix T is given as:

This matrix is also illustrated in _{1 }relative to W_{2 }and W_{3}.

Before considering the possibilities of achieving even further reductions, it is useful consider the matrix shown in

With further regard to ^{−1}=T^{30}, T^{2}=T^{29}, T^{−3}=T^{28}, etc. This fact is relevant in comparing symbol column H_{Q }with Equation (6). Similar statements are applicable to the other symbol columns.

With particular attention focused on matrix row # 2 from ^{14}. Again, because of the periodic nature of the powers of T, it is possible to have such matrices range all the way up to T^{m }where m=2^{b}−2. If b=5, m=30. Suppose that H_{max }is a parity check matrix having the maximal number of symbol columns. This affords the opportunity for even greater reductions. In particular, in practical applications, the code length is usually less than the maximum possible value for a given code class. In particular, the number of symbols required for the code may be less than N. In this case, it is possible to select a subset of symbol columns from H_{max }to form the desired parity check matrix. In order to minimize the number of Exclusive-OR gates required in a hardware coding or decoding implementation, those symbol columns are chosen which have a minimal weight. This leads to yet a new method for parity check matrix construction. In particular, the first step of this method involves the application of the matrix multiplication and weight selection algorithm described above but now with m having its maximum value 2^{b}−2. Then from a modified H_{max }the desired number of symbol columns are selected so as to form a subset having the smallest total weight to use as a reduced parity check matrix.

These principles can be applied to the example presented above in which b=5, N=18 and p(x)=1+x^{2}+x^{5}. This is the same set of parameters used in _{max}), an even further reduced parity check matrix is found. This matrix is shown in

Although the considerations provided above have generally only been directed to codes having three check symbols, the methods of the present invention can be applied to constructing a byte organized error correcting code with minimum weight symbol columns in its parity check matrix for any number of check symbols. In particular, it is seen in ^{b}−2 such powers of T to consider. Nonetheless, in the more general case, it is possible to explore the results obtained by using all possible non-singular b×b matrices. While this problem becomes much more difficult with higher numbers of symbol bits b, the problem is nonetheless tractable when b=3, 4 or 5. Nonetheless, it is seen that since each symbol column can be manipulated independently of the others, the multiplication and simplification operations are nonetheless straight forward and do not grow rapidly with the number of symbol columns or with the number of check symbols.

Attention is now specifically directed to the advantages of the present invention which are applicable to decoding. Certainly, it is seen that for encoders, the smaller the number of 1's in the parity check matrix the smaller will be the amount of hardware required to implement it since the number of two-way Exclusive-OR gates required with the i^{th }check bit is equal to the number of 1's in the i^{th }row of the binary matrix, minus one. With respect to decoders, the situation is in general not as simple. However, because of simplifications brought about by the presence of an identity matrix as one of the submatrices in each symbol column, it is possible to exploit certain code structures to provide decoders having a minimal number of Exclusive-OR gates. Furthermore, such decoders are readily designable using automated circuit layout algorithms and methods.

In particular, suppose that E is the b bit error pattern for a symbol. The positions of errors are indicated by the presence of 1's in the E vector. Suppose that the j^{th }symbol of a code word contains E as the error pattern for that symbol. The error syndrome of E is equal to the product of the j^{th }column of the parity check matrix H and the vector E. The syndrome S is given as Hy, where y is the received code word vector. Thus, if the code vector x is sent and the vector y is received then x=y−E (or y=x+E). However, it is known that Hy=HE because x is a code word for which Hx=0. In particular, in this situation in which there are three check symbols, the syndrome S can be partitioned into three subvectors of b bits each. This partitioning is particularly illustrated in _{1 }S_{2 }S_{3}). Suppose further that the three submatrix components of the j^{th }symbol column of H are denoted by H_{1j}, H_{2j }and H_{3j}. This organization is particularly illustrated in

*S* _{1} *=H* _{1j} *·Ē* (8a)

*S* _{2} *=H* _{2j} *·Ē* (8b)

*S* _{3} *=H* _{3j} *·Ē* (8c)

In the above, S_{1}, S_{2 }and S_{3 }and Ē are considered as column vectors and Ē is the b bit error pattern for symbol j from the n bit long E vector. For each and every symbol column j one of the submatrices H_{ij}, i=1, 2, 3, represents an identity matrix and accordingly one of either S_{1}, S_{2 }or S_{3 }is the error pattern Ē for symbol j. Which one of the S_{i }is the identity matrix depends upon j, the particular symbol column being considered. However, this is known from the structure of the original parity matrix H which is used to encode the vector which has been received. Thus, the error pattern Ē can be readily identified as one of the three subvector components of S. Once Ē has been identified, the symbol error status is determined by checking that all three congruence relations of Equation (8) are true.

For example, consider the code defined by the parity check matrix in _{2 }since H_{12 }is an identity matrix in the first symbol column. To verify if the errors are really in the first symbol position, it is necessary to check that S_{1}=T^{18}·S_{2 }and that S_{3}=T^{5}·S_{2}. While this computation must be carried out for each symbol column separately, it is nonetheless seen that there are relatively easy ways of carrying out the computation and the necessary comparison.

Attention is now specifically directed to decoding methods and apparatus for code words constructed in accordance with certain specific embodiments of the present invention. In particular, **6**D and **6**E illustrate more detailed aspects of the system shown in

With respect to **1** through M. Each such symbol error detection unit **10** essentially performs a binary matrix multiplication operation and a comparison operation. The output of functional block **10** is essentially an enable signal supplied to a corresponding one of the bit error detection and correction units **20**. Correction units **20** receive an enable signal from a corresponding symbol error detection unit **10** and each unit **20** also receives a specific subset of syndrome data bits. In each case, the subset of syndrome bits received matches the syndrome subvector having the same relative position in the syndrome as the identity matrix has in the associated symbol column from H. Additionally, for each separate symbol, bit error detection and correction units **20** receive the symbol data bits themselves. Functional block **20** therefore typically receives b syndrome bits and b symbol data bits. Additionally, functional block **20** receives the enable signal from symbol error detection unit **10**. This enable signal is also supplied to uncorrectable error detection unit **30**. Each of these units is more particularly described below in the discussions relating to

Attention is now specifically directed to symbol error detection unit **10**. Such a unit is shown in more detail in ^{th }symbol. In the case shown, the syndrome is partitioned into three groups of b bits each. Two matrix vector product operations are performed, one of which computes the product H_{ij}·S_{k}, where i is not equal to k and where k depends on j. The other matrix product computed by unit **10** is the matrix vector product H_{1j}·S_{k}, where 1 is not equal to k, which depends on j. It is noted that each symbol error detector **10** is associated with its own unique value of k. In particular, k is the identifier of the matrix row in the symbol column which includes the identity matrix. For the first symbol column in the parity check matrix shown in _{11}·S_{2 }and H_{31}·S_{2}. More particularly, if we also denote the syndrome vector S as (s_{1 }s_{2 }. . . s_{15 }then symbol column multiplier **12** is actually seen to carry out the following two matrix operations given by Equations (9a) and (9b) below:

These last two equations, particularly when expressed in terms of Exclusive-OR or modulo 2 operations, clearly indicate that circuit complexity is directly proportional to the number of 1's in the parity check matrix. These equations also directly specify circuitry which implements the matrix products indicated. Bear in mind that Equations (9a) and (9b) represent the situation only for the matrix of **1**) for which it is seen from _{1 }and S_{3 }respectively. This operation is carried out by means of comparators **13** and **14** respectively as seen in **10** however, the comparison results are supplied to NOR-gate **15** which provides the desired enable signal to a bit error detection and correction unit and to uncorrectable error indicator **30**.

Bit error detection and correction units are more specifically illustrated in _{k}, namely bits (S_{k1 }S_{k2 }. . . S_{kb}) are provided to AND gates **21**. AND gates **21** receive the enable signal from symbol error detection unit **10** so as to permit correction of bits for that symbol position only. Thus, the error pattern represented by syndrome subvector S_{k }is supplied under appropriate conditions directly to Exclusive-OR gates **22**. These Exclusive-OR gates receive symbol data bits which are corrected by applying the error pattern (syndrome subvector) to Exclusive-OR gates **22** which operate as conditional inverters. Thus, where necessary, received data bits are inverted in accordance with the error pattern. This is made possible by the presence of an identity matrix in each of the symbol columns of H. Thus, bit error detection and correction unit **20** operates to produced corrected symbol bits for symbol j, as shown.

Attention is next directed to uncorrectable error indicator **30** which receives enable signals from symbol error detection units **10** which are thus supplied to multiple input NOR-gate **31**. The output of this NOR-gate is directed to AND gate **32** which passes this signal whenever a non-zero syndrome is detected. See

The relation between decoding hardware and the submatrices of a parity check matrix H is particularly illustrated in **10** and bit error detection and correction unit **20** for the first symbol position (j=1) associated with the parity check matrix of **10** of **10** and Exclusive-OR gates **16**.**1** through **16**.**5** which carry out the matrix multiplication operations shown in Equation (9a). The matrix-vector multiplication operation (actually a sequence of modulo 2 summations) is carried out using the syndrome bits supplied to Exclusive-OR gates **16**.**1** through **16**.**5** which are shown having both short lead lines. The short lead lines in **17** complementing this inverse). Thus, in order to compare the matrix vector product H_{11}·S_{2 }with the syndrome vector S_{1 }it is seen that it is only necessary to supply the syndrome subvector bit positions s_{1 }through s_{5 }to Exclusive-OR gates **16**.**1** through **16**.**5** respectively. (Be sure to distinguish S_{1 }a b bit syndrome subvector from S_{1 }the first syndrome bit.) This is shown using the longer input lead lines. Thus the multiplication and comparison operations indicated in **16**.**6** through **16**.**10** effectively carry out the matrix multiplication indicated in Equation (9b) namely the matrix-vector product H_{31}·S_{2}. This result is effectively compared with syndrome subvector S_{3 }by supplying syndrome bit positions s_{11 }through s_{15 }to Exclusive-OR gates **16**.**6** through **16**.**10** respectively. The Exclusive-OR gate outputs (gates **16**.**1** through **16**.**10**) are supplied to NOR-gate **17** which provides the desired enable signal to AND gates **21**.**1** through **21**.**5** in bit error detection and correction unit **20**. These AND gates also receive the error vector pattern which for symbol position **1** as represented by syndrome subvector S_{2}=(s_{6 }s_{7 }s_{8 }s_{9 }s_{10}) or using the notation shown in _{2}=(S_{21 }S_{22 }S_{23 }S_{24 }S_{25}). The error pattern is supplied to AND gates **21**.**1** through **21**.**5** respectively which are enabled by the output from multiple input NOR-gate **17**. The outputs of AND gates **21**.**1** through **21**.**5** are, as described above, supplied to Exclusive-OR gates **22**.**1** through **22**.**5** which act as conditional inverters for received data symbol bits D_{1 }through D_{5}. Thus, the output of bit error and correction unit **20** is the corrected symbol bits for symbol position No. 1. Similar decoding circuitry is provided for each of the other symbol positions. Furthermore, an uncorrectable error indication is provided through the operation of functional block **30** as shown in

Next is considered the specific relation between the parity check matrix shown in _{10}. However, other simplifications result from the elimination of other inputs to Exclusive-OR gates **16**.**1** through **16**.**10**. Moreover, fewer AND gates and Exclusive-OR gates are required in bit error detection and correction unit **20**.

One of the advantages of the circuitry shown in

It is further recognized by those skilled in the art of error correction code design that it is possible to derive other codes from a given parity check matrix. As indicated above, some of these codes are obtained by deleting columns of a parity check matrix. Accordingly, the present invention is also seen to be directed to coding systems which are derivable from the parity check matrix modifications of the present invention. In this regard, it is noted that in the process of deleting binary columns, particularly as seen in the transition from

Accordingly, it is seen that the present applicant has provided a method and apparatus for coding and decoding which results in low cost circuitry and improved decoding times. It is further seen that the present applicant has employed hitherto unknown techniques to manipulate and improve the performance of a large class of codes which are well known and which have highly useful error correcting properties. It is further seen that the applicant has implemented otherwise complex matrix-vector product and comparison operations in a single layer of circuitry, thus enhancing speed of operation while nonetheless providing code manipulation methods for insuring that the present codes can be implemented using minimal hardware.

While the invention has been described in detail herein in accordance with certain preferred embodiments thereof, many modifications and changes therein may be effected by those skilled in the art. Accordingly, it is intended by the appended claims to cover all such modifications and changes as fall within the true spirit and scope of the invention.

Patent Citations

Cited Patent | Filing date | Publication date | Applicant | Title |
---|---|---|---|---|

US3601798 | Feb 3, 1970 | Aug 24, 1971 | Ibm | Error correcting and detecting systems |

US4358848 | Nov 14, 1980 | Nov 9, 1982 | International Business Machines Corporation | Dual function ECC system with block check byte |

EP0300139A2 * | Apr 19, 1988 | Jan 25, 1989 | International Business Machines Corporation | Error correcting code for B-bit-per-chip memory with reduced redundancy |

JPS60178717A * | Title not available |

Non-Patent Citations

Reference | ||
---|---|---|

1 | "A High Speed Parallel Encoder/Decoder for b-Adjacent Error-Checking Codes", Computer Conference, 1978, pp. 203-207. | |

2 | "Digital Systems and Hardware/Firmware Algorithms", Milos D. Ercegovac and Tomas Lang, 1985, pp. 126-127. | |

3 | "Fault-Tolerant Memory Design in the IBM Application System/400", C.L. Chen and L.E. Grosbach, IEEE 1991, pp. 393-400. | |

4 | "Single Byte Error Correcting-Double Byte Error Detecting Codes for Memory Systems", Shiego Kaneda and Eiji Fujiwara, IEEE Transactions on Computers, vol. C-31, No. 7, Jul. 1982, pp. 596-602. | |

5 | * | Bhatt et al., A High Speed Parallel Encoder/Decoder for b-Adjacent Error-Checking Codes, 3rd USA-Japan Computer Conference, 1978, pp. 203-207. |

6 | * | Bossen, D., "b-Adjacent Error Correction", IBM Journal of Research & Development, Jul. 1970, pp. 402-408. |

7 | * | Glick, E., et al., "Single-Error Correction, Double-Error Detection Code Utilizing Minimum Circuitry", IBM Technical Disclosure Bulletin, vol. 15, No. 1, Jun. 1972, pp. 130-134. |

8 | Hsiao, M.Y., A Class of Optimal Minimum Odd-weight-column SEC-DED Codes, Jul. 1970, IBM Journal of Research and Development. | |

9 | IBM Journal of Research and Development, Jul. 1970, M. Y. Hsiao, pp. 395-401, A Class of Optimal Minimum Odd-weight-column SEC-DED Codes. | |

10 | * | No Author, "Dual-Mode Error Correction and Error Detection", IBM Technical Disclosure Bulletin, vol. 28, No. 1, Jun. 1985, pp. 55-58. |

Referenced by

Citing Patent | Filing date | Publication date | Applicant | Title |
---|---|---|---|---|

US7721140 * | Jan 2, 2007 | May 18, 2010 | International Business Machines Corporation | Systems and methods for improving serviceability of a memory system |

US7721177 * | May 12, 2006 | May 18, 2010 | Infineon Technologies Ag | Device and method for determining a position of a bit error in a bit sequence |

US8065584 * | May 21, 2007 | Nov 22, 2011 | Hewlett-Packard Development Company, L.P. | Transmitting data words |

US8327105 | Feb 16, 2012 | Dec 4, 2012 | International Business Machines Corporation | Providing frame start indication in a memory system having indeterminate read data latency |

US8381064 | Jun 30, 2010 | Feb 19, 2013 | International Business Machines Corporation | High density high reliability memory module with power gating and a fault tolerant address and command bus |

US8419547 | Apr 16, 2013 | Wms Gaming, Inc. | Iterative XOR-matrix forward error correction for gaming | |

US8495328 | Feb 16, 2012 | Jul 23, 2013 | International Business Machines Corporation | Providing frame start indication in a memory system having indeterminate read data latency |

US8589769 | Sep 7, 2007 | Nov 19, 2013 | International Business Machines Corporation | System, method and storage medium for providing fault detection and correction in a memory subsystem |

US8966351 * | Dec 6, 2012 | Feb 24, 2015 | Kabushiki Kaisha Toshiba | Encoding apparatus, encoding method and semiconductor memory system |

US20060282756 * | May 12, 2006 | Dec 14, 2006 | Infineon Technologies Ag | Device and method for determining a position of a bit error in a bit sequence |

US20070198904 * | Oct 24, 2006 | Aug 23, 2007 | Kabushiki Kaisha Toshiba | Error correction processing apparatus and error correction processing method |

US20070220393 * | May 21, 2007 | Sep 20, 2007 | Hewlett-Packard Development Company, L.P. | Transmitting data words |

US20080162991 * | Jan 2, 2007 | Jul 3, 2008 | International Business Machines Corporation | Systems and methods for improving serviceability of a memory system |

US20130151921 * | Dec 6, 2012 | Jun 13, 2013 | Hironori Uchikawa | Encoding apparatus, encoding method and semiconductor memory system |

Classifications

U.S. Classification | 714/752 |

International Classification | H03M13/13, G06F11/10, H03M13/00, H03M13/11 |

Cooperative Classification | H03M13/13 |

European Classification | H03M13/13 |

Legal Events

Date | Code | Event | Description |
---|---|---|---|

Sep 17, 2012 | REMI | Maintenance fee reminder mailed | |

Feb 3, 2013 | LAPS | Lapse for failure to pay maintenance fees | |

Mar 26, 2013 | FP | Expired due to failure to pay maintenance fee | Effective date: 20130203 |

Rotate